1609685371 / adaptive-wing-loss-for-robust-face-alignment-via-heatmap-regression

Adaptive-Wing-Loss-for-Robust-Face-Alignment-via-Heatmap-Regression

Adaptive-Wing-Loss-for-Robust-Face-Alignment-via-Heatmap-Regression

❗ongoing repo

Pytorch implementation of paper. official implementation can be found at official.

blog post about the paper(korean) can be found here.

📝 TODO

• prototype
• albumentation data augmentation
• evalutaion on 300W + data augmentation
• performance tuning
• dependency check
• provide pretrained weight
• apply different model (such as DLA, Unet)
• apply similar loss (such as Focal-loss)
• apply Integral regression moduel (AWing + Integral)

Prerequisites

• Python 3.6 +
• Pytorch 1.1.0
• Scipy 0.19.1
• cv2 3.3.0

Usage

First, download dataset(Currently 300W supported).

300W link

1. download [part1] ~ [part2]
2. locate 300W images, pts files according to this structure

data

|-- 300W
|   |-- 01_Indoor
|   |-- 02_Ourdoor

To train a model with downloaded dataset:

$ python train.py

To test single image result:

$ python detect.py

Model overview

more detail about model

loss function design

AWing → (lossMatrix) → Loss_weighted

evalutaion

evalutaion on 300W testing dataset

evaluation result will soon be updated

method	NME	FR(10)
the paper	3.07	X
this repo	x	0.8

Reference

• CoordConv
• Stacked Hourglass
• AdaptiveWingLoss